Presented by Dr Niloy Ganguly Department of Computer

Presented by: Dr. Niloy Ganguly Department of Computer Science, IIT Kharagpur. Co-authors – Ujjwal Sarkar, Subrata Nandi 1

Outline v Introduction v Motivation v Simulation Environment v Algorithm v Result Analysis v Conclusion v Reference 2

Peer-to-Peer Architecture v v A cooperative resource sharing environment. Efficient sharing of computer resources and services by direct exchange between systems. Virtual overlay on the top of existing network with own routing mechanism. Structured or unstructured topology. Characteristics: No fixed client or server � Decentralization � Dynamic � Self organizing � Anonymous � 3

Gnutella v Decentralized, unstructured, content sharing P 2 P network. v Open system architecture. v Uses flooding to locate resource. v Popular contents are replicated. v Power law topology. Courtesy: Lua, Crowcroft, IEEE Comm. 04 4

Observations on Gnutella 1. Free Riding 2. Security Threats in P 2 P Network 3. Poor search scalability 5

Observations 1)Free Riding Ø Manifestation of tragedy of commons § 70% of users share no file. § 1% of hosts answer nearly 50% of all queries. § 25% users account for 99% of all queries. Ø Indicates peers are heterogeneous entities. Ø The system goal differs from individual goal. ( Ref: E. Adar, “Free Riding on Gnutella”, First Monday, September 2000) ( Ref: M. Ripeanu, “Peer-to-peer architecture case study: Gnutella network”, P 2 P Computing ’ 01) v Solution: To provide incentive to upload files. 6

2) Security Threats in P 2 P Network v Fake Content distribution (Ref: J. SchÄafer, “P 2 P networks security”, ICIMP '08 ) v Malicious File Upload e. g. VBS. Gnutella. worm (Ref: Ernesto Damiani, “A reputation-based approach for choosing reliable resources in P 2 P N/W”, CCS '02) v White washing (Ref: Michal Feldman, “Free-riding and white washing in P 2 P system”, PINS '04) Solution: Download from trusted source 7

Trust Management Schemes v Trust is probability that resource provider will provide authentic files. v Computed based on transaction history. v Direct trust and Recommendation trust. Challenges: Decentralized and scalable. • Cope up with transient nature of P 2 P. • Robust against various threat models. Types: 1) Centralized Reputation based trust management § used in Ebay, amazon. com 2) Distributed Reputation based scheme § used in P 2 P network. 8

Reputation Based Trust Management (unstructured network) v Allows resource requester to compute trust rating of the resource provider. v. Types § Gossip based e. g. XREP, Eigen Trust § via Topology Adaptation e. g. APT, RC-ATP. A natural choice for unstructured topology v Limitation of existing trust management scheme Heavy weight: High computational cost, message and storage overhead. § Lack effective mechanism to disseminate trust information. § Presence of dynamicity (churning) is not taken in amount. § 9

3) Poor search scalability v Flooding § v Random walker § v BFS with limited TTL A blind search via Topology adaptation Semantic community § efficient § Iterative deepening § 10

Illustration P 2 P overlay Trust management Semantic Communities 11

Motivation v Both search quality and efficiency equally important. v Existing trust management schemes are heavyweight. v No work carried out to use topology adaptation to combat inauthentic downloading as well as to improve search scalability. v To incorporate incentives and punishment mechanism to combat free riding and fake content distribution. Nodes get central position as a reward. Trust aware community is proposed to address above issues 12

What is Trust Aware Community? v An overlay network of trusted peers. v Neighbors are selected based on trust and content similarity. v Evolving search strategy. 13

What is Trust Aware Community? v An overlay network of trusted peers. v Neighbors are selected based on trust and content similarity. v Evolving search strategy. P 2 P overlay Trust aware community 14

Simulation Environment 1) Network Topology & Load v Power law graph. Connectivity link and Community links. Increase in degree is in constrained by initial degree. v v 2)Content Distribution Model v Each file is represented by a tuple (c, r) where c content category, r rank of file. v More popular categories are more replicated. v Within a category popular files are more replicated. v Follows zipf’s law. (Ref. Kamvar, “Simulating a P 2 P file-sharing network”, P 2 P and Grid Computing ’ 02) 15

contd. 3) Query Model v v Use Poison distribution to calculate number of queries each peer issues. A peer issues queries for files not present in its own categories. 4)Threat Model v Model A: Malicious peers provides good files probabilistically. v Model B: Malicious peers provides fake file only when it gains sufficient community edges. 16

Outline Of Algorithm v Network learn trust through search. v Five basic modules 1. Search/ Forward 2. Response selection and download 3. Update trust 4. Check trust 5. Rewire topology 17

Trust Aware Topology Algorithm 1)Search/Forward v Uses Directed BFS which evolves to DFS as network connectivity increases using parameter v Queries are disseminated through trusted neighbors, among trusted neighbors matching community members are preferred. v Queries forwarded by malicious peers are dropped. 2) Response selection: Response are sorted based on trust rating of source peers. v If trust rating of source is not available in local db, recommendation is sought via trust query. v 18

BF Search tree illustrating search initiated by peer 1. v Uses TTL limited modified BFS which evolves to DFS as network connectivity increases. v Queries are disseminated through trusted neighbors, among trusted neighbors matching community members are preferred. v Queries forwarded by mal. peers are dropped. 19

CONTD. 3) Topology adaptation: v v After successful download a peer probabilistically attempt to form link with resource provider. Requires approval of resource provider. v If trust rating of source is negative, existing community edge is removed unconditionally. v Controlled by parameter edge limit and degree of rewiring.

Figure illustrates topology adaptation. Mal. nodes are shaded in grey. 21

CONTD. 4) Trust Updating: v v v LRU structure used to remember past transactions with other peers. After each download of file from peer ‘j’ , peer ‘i’ changes by +/ - 1, where be # of successful transaction. Normalized value of be trust score of peer ‘j’ as per peer ‘i’ local history. 5) Trust Query: v A TTL limited DFS to seek recommendation from neighbors. v Query is propagated at each hop through a trusted neighbor v Uses iterative deepening. 22

Metrics 1) Search Qo. S related metrics: v Attempt Ratio (AR): It is the probability of downloading a file in the first attempt. Let P be the total number of attempts to download an authentic file, then attempt ratio is defined as AR=1/P*100 or zero, if it fails to download authentic file. v Effective Attempt Ratio (EAR): Let P(i) be the total number of attempts made by peer ‘i’ to download an authentic file. Then where M and N be the number of good and malicious peers v Query miss ratio (QMR): Fraction of total search failures in a single generation. 23

2) Topology related metrics v Largest connected component (LCC): Fraction of total peers in largest connected component sharing a particular content category. v Relative increase in connectivity (RIC): where N be total # of peers Simulation parameters 24

Fig illustrating search quality Comparison with an equivalent network 25

Fig. illustrating Search efficiency Performance under node churn 26

Figures Illustrating goodness of community formation 27

Attempt Ratio Relative increase in connectivity Performance of free rider 28

Comparison with existing techniques Eigen trust: § Fraction of response is high for good peers when percentage of malicious peers is ~80. Trust aware topology can withstand up to 60 %. § Eigen trust is computationally intensive. § Eigen value converges only in static network and suffers from Byzantine consensus problem. APT/RC-ATP: § Trust aware topology is scalable , but RC-ATP not. § Fraction of authentic response is 100 % for good peers. with 10 % malicious peers. § RC-ATP not evaluated with higher percentage of malicious peers. § Use flooding to locate files. Trust aware community use evolving search. 29

Conclusion v Trust aware community combats fake download, free riding and poor search scalability. v It is scalable and light weight. v Incorporates incentives and punishment mechanism. v White washing is not considered. v Not tested in real network data. 30

References § Ernesto Damiani, De Capitani di Vimercati, Stefano Paraboschi. “A Reputation based Approach for Choosing Reliable Resources Peer to Peer Networks”, Proceedings of the 9 th ACM conference on Computer and communications security, 2002. § Eytan Adar and Bernardo A. Huberman. “Free Riding on Gnutella”, First Monday 5, October 2000. § Michal Feldman, Christos Papadimitriou, John Chuang, Ion Stoica. “Free. Riding and Whitewashing in Peer-to-Peer Systems”, SIGCOMM-04 Workshop, August-September, 2004. § Matei Ripeanu. “Peer-to-peer architecture case study: Gnutella Network”, Proceedings of First International Conference on Peer-to-Peer Computing, 2001. § A. Abdul-Rahman and Stephen Hailes. “A Distributed Trust Model”, Proceedings of the 1997 workshop on New security paradigms, Pages: 48 -60, 1998. § Sepandar D. Kamvar, Mario T. Schlosser, Hector Garcia-molina. “The Eigen Trust Algorithm for Reputation Management in P 2 P Networks”, In Proceedings of the Twelfth International World Wide Web Conference, 2003. 31

Contd. § Tyson Condie, Sepandar D. Kamvar, Hector Garcia-Molina. “Adaptive Peer-To-Peer Topologies”, Proceedings of the Fourth International Conference on Peer-to-Peer Computing, 2004. § Huirong Tian, Shihong Zou, Wendong Wang, Shiduan Cheng. “Constructing efficient peer-to-peer overlay topologies by adaptive connection establishment”, Computer Communications Volume 29, Issue 17, 8 November 2006. § Kevin Walsh, Emin Gun Sirer. “Fighting Peer-to-Peer SPAM and Decoys with Object Reputation”, P 2 PECON Workshop, Philadelphia, Pennsylvania, USA, August 2005. § Kunwadee Sripanidkulchai, Bruce Maggs, Hui Zhang. “Efficient Content Location Using Interest-Based Locality in Peer-to-Peer Systems”, INFOCOM 2003. § Vicent Cholvi, Pascal Felber. “Efficient Search in Unstructured Peer-to. Peer Networks”, European Transactions on Telecommunications: Special Issue on P 2 P Networking and P 2 P Services, 2004. § Tathagata Das, Subrata Nandiy and Niloy Ganguly. “Community Formation and Search in P 2 P: A Robust and Self-Adjusting Algorithm”, IAMCOM 2009, held with COMSNET 2009. 32

Contd. § Liangmin Guo, Shoubao Yang, Leitao Guo, Kai Shen, Weina Lu. “Trustaware Adaptive P 2 P Overlay Topology Based on Superpeer-partition”, Proceedings of the Sixth International Conference on Grid and Cooperative Computing, 2007. § Mario T. Schlosser, Tyson E. Condie, Ar D. Kamvar. “Simulating a P 2 P file -sharing network”, First Workshop on Semantics in P 2 P and Grid Computing, 2002. § A. Crespo, H. Garcia-Molina. “Semantic Overlay Networks for P 2 P Systems”, Technical report, Computer Science Department, Stanford University, 2002. § S. Saroiu, P. K. Gummadi, S. D. Gribble. A. “A Measurement Study of Peer-to-Peer File Sharing Systems”, Proceedings of Multimedia Computing and Networking, 2002. 33

Dynamics On and Of Complex Networks Applications to Biology, Computer Science, and the Social Sciences Ganguly, Niloy; Deutsch, Andreas; Mukherjee, Animesh (Eds. ) A Birkhäuser book Workshop – 23 rd September, Warwick

Thank you

Thank You 36